Eternis-Forecaster 8B: Activation Probes Beat Chains of Thought and Save Up to 47% of Tokens

On July 10, 2026, researchers published a paper on arXiv showing that the forecasts of the Eternis-Forecaster 8B language model are "locked in" before chain-of-thought reasoning even begins. Probes of intermediate activations also calibrate the model's confidence more accurately and predict changes in its behavior in 84% of cases, even when the text reasoning chain hides those changes.

What the researchers found

The team worked with Eternis-Forecaster 8B on the OpenForesight benchmark, a specialized platform for probabilistic forecasting. They trained probes with representation pooling on the model's intermediate activations, and these showed substantially better calibration than the model's own built-in confidence. The results replicated on GLM-4.7-Flash and GLM-4.5-Air, which confirms that the approach transfers across models.

Key facts:

• Models tested: Eternis-Forecaster 8B, GLM-4.7-Flash, GLM-4.5-Air
• Benchmark: OpenForesight (probabilistic forecasting)
• Accuracy in predicting the direction of changes: 84% of cases
• Token savings from routing: 30–47% without loss of accuracy
• Method: representation pooling on intermediate transformer layers

Why chains of thought should not be trusted blindly

The authors tested the "faithfulness" of CoT in two ways: removing influential sources from the prompt and injecting distracting information. The pattern held up. Removing a key source often changed the final forecast, yet the text reasoning stayed untouched: the model never mentioned the vanished source and behaved as if it had never been in the context.

The probes behaved fundamentally differently. Their activations reliably tracked behavioral shifts and predicted the direction of changes in 84% of cases, including when the CoT completely concealed the perturbation. That makes activation probes a considerably more reliable auditing tool than the CoT itself.

"Internal representations reveal what the model actually processes,

not what it chooses to report" is the conclusion the authors draw from the full body of their experiments.

The forecast is fixed before reasoning begins

The most practically significant result: the model's answers are already "locked in" before the reasoning chain starts. A single preliminary pass without generating CoT is enough to accurately recover both the final answer and the model's confidence in it.

This has a direct practical use: smart query routing based on the spread of the preset answer distribution. Where the answer is already clear at the activation level, expensive reasoning generation can be skipped, saving 30% to 47% of tokens without loss of accuracy in the final forecasts.

What this means

Probes of internal representations give direct access to what the model actually processes, bypassing text reasoning, which is not obliged to reflect the real causes of an output. For forecasting systems, this means a practical toolkit: confidence calibration, CoT faithfulness auditing and efficient query routing in a single method. The authors see the approach as applicable not only to forecasters but to a broader class of reasoning models. As LLMs are increasingly used in decision-making, this is an important step toward systems whose internal beliefs can be measured independently of what the model says.

Frequently asked questions

What are activation probes and why are they more accurate than

CoT?

Activation probes are small classifiers trained on the intermediate states of a neural network. They read the model's "internal signal" directly, bypassing text output. In this study's experiments, probes with representation pooling predicted Eternis-Forecaster 8B's final forecast and the direction of its changes more accurately than text CoT reasoning.

Does the method apply to other models and tasks?

The authors replicated the calibration improvement on GLM-4.7-Flash and GLM-4.5-Air, confirming that the approach transfers. The main benchmark, OpenForesight, is specific to probabilistic forecasting; applicability to other tasks and domains requires separate verification.